Natural language understanding technologies involve the identification of the intended semantic from the multiple possible semantics. This means that the system needs a contextual lexicon of the language, with a suitable ontology, in order to provide to machines a mental model of the surrounding world (VirtualBrains) and be able to assess the provided data to extract his meaning, in several languages and without prior training.
Today, when people talk about artificial intelligence, frequently topics like machine learning or deep learning are discussed. Mainly based on mathematics or statistics, these forms of learning for a machine are supervised or unsupervised. Supervised learning is when the machine needs to find data which are matching a given pattern while unsupervised learning is when the machine must analyze data to find a structure (like for speech or face recognition). To do this, data scientists are needed. Machine learning became a technology, driven by mathematicians or statisticians.
You remember those three wise monkeys? Mizaru covered his eyes, Kikazaru covered his ears and Iwazaru covered his mouth. They did this to see no evil, hear no evil and speak no evil. Indeed, if one cannot see, he cannot acquire any visual information. If one cannot hear, he cannot process any audible stimulation. If no information is captured, nothing can be learned, nothing can be reported.
For Mizaru to learn from what he sees he needs to have his eyes wide open. Learning is the ability to acquire knowledge. Acquisition of knowledge relies on perception which is the organization, identification, and interpretation of sensory information to represent and understand the presented information, or the environment.
The sensory system is the input system which takes physical or chemical stimulation and transform it into signals going through the nervous system. Perception is not only the passive receipt of these signals, but it’s also shaped by the recipient’s learning, memory, expectation, and attention. Perception can be split into two processes, (1) processing the sensory input, which transforms this low-level information to higher-level information (e.g., extracts shapes for object recognition), (2) processing which relates to a person’s concepts and expectations (or knowledge) and selective mechanisms (attention) that influence perception. Perception depends on complex functions of the nervous system, but subjectively seems mostly effortless because this processing happens outside conscious awareness.
This means that to perceive, an input system is needed, and this input system will transform the incoming physical or chemical stimulation into a mental information so that this mental information can be processed further. And this mental information could be perceived as not useful and forgotten or could be further decoded and used by other higher mental processes or by cognition. But why? Just because as it is said above, attention will influence perception. If one’s expectation is to collect more information on a given topic, he will pay attention to incoming stimulations.
Let’s take a very simple example. It is lunch time; I have an appointment with my friend. We will eat together. I’m waiting somewhere city center and around me are many people, noises and cars. I look around to see where my friend is. How many stimulations I have right now? All the pictures (buildings, cars, people moving), all the sounds (people talking, engine noises, horns) … Will I process all these stimulations? Yes, of course. I can see all the people, all the cars, I can hear all the noises, the baby who is crying and this motorcycle engine sound. Do I really take care of these stimulations? Well, not really because I’m focused just on one thing: finding my friend. This means that my attention is more focused on faces and if my friend is blond, more on blond people, if my friend is a girl, more on blond girls. So, I still process all incoming information, but I’m focused, I pay attention on stimulations like human faces to find my friend. I may hear an engine, but I do not process this stimulation further since it is not relevant to me in this context.
Transforming the incoming physical or chemical stimulation into a mental information so that recognition can take place requires that the mental information to be found is already known or that the intention to find something which could satisfies that purpose is found. In the example above, I can recognize my friend because I know my friend. It is not possible to recognize someone or something that I never saw before! I could also recognize someone as fitting the purpose of my search since this someone is satisfying me or filling my intentions. It is the nearest object among all available who could satisfy me.
The hard part of the job when the machine is learning is not to store the information but to recognize among all incoming data those who should be considered for being processed. Those pattern recognition algorithms are computing a probable match between data being processed and a specific reference. Those machine learning algorithms are good at recognizing an information but not at making sense of it. Those algorithms are good at classifying information but not at categorizing them or in other words are not able to build without human intervention ontologies made of multiple semantic fields.
To make it very clear, let’s take an example and let’s dig into the details of what we just said.
Let’s say you are deploying an artificial intelligence that will use unsupervised machine learning algorithm to make sense of the content provided in natural language within business related verbatims in the banking sector. At the end of this processing, some words will appear as being more frequent than others, like the word “credit”.
But “credit” is just a word. Since people, can use other words to express the same idea, they could use synonyms instead of the word “credit”. Since I do have access to a lexicon containing all the available synonyms for a given word, I’m able to retrieve them and once this is done, I will have a group of words related to the word “credit”. I will have a semantic field related to the word “credit”. This approach is wrong, completely wrong.
Let’s do it.
The word “credit” can be used to express an approval with following synonyms: recognition, ovation, salutation, remembrance, commendation, approval, salute, standing ovation or commemoration. It can be used to express money available for a client to borrow with following synonyms: export credit, cheap money, assets, credit line, personal credit line, line of credit, line, letter of credit, import credit, bank line or commercial credit. It can also be used to give someone credit for something with following synonyms: ascribe, attribute, impute… There are still other possible meanings.
This would mean that the semantic field related to the word “credit” would be made of following words: recognition, ovation, salutation, remembrance, commendation, approval, salute, standing ovation, commemoration, export credit, cheap money, assets, credit line, personal credit line, line of credit, line, letter of credit, import credit, bank line, commercial credit, ascribe, attribute and impute.
Since the artificial intelligent system should be able to recognize words, it will most probably do a correct job by recognizing within a given verbatim those words which are belonging to the semantic field of the word “credit”. So, if within a verbatim I have sentences such as “I give my approval to move to San Francisco” or another sentence such as “My computer program is accepting following attributes”, both sentences do contain words which do belong to the semantic field of the word “credit” and those two sentences should be tagged as such!
You can see that building an artificial intelligent system which purpose would be to understand the content of a verbatim cannot be done like this. This is making no sense.
There is something missing.
For this exact purpose, Percipion is knowledge based. It already has a contextual lexicon. But having this lexicon is not enough to make sense of this information or to understand it. To be able to do this, this knowledge has to be organized like it is organized within our long-term memory.